Varto
Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>

Over 35,000 apps built with Grok Build have sensitive data exposed. SpaceXSI, which created Grok, is among those affected

Apps built with the AI ​​tool Grok Build were found to contain confidential business plans, employee personal data, and even patient medical records. Among the victims was SpaceXSI, which owns Grok. All it took was a link to access this information.

Leave a comment
Add to Google
Over 35,000 apps built with Grok Build have sensitive data exposed. SpaceXSI, which created Grok, is among those affected

Apps built with the AI ​​tool Grok Build were found to contain confidential business plans, employee personal data, and even patient medical records. Among the victims was SpaceXSI, which owns Grok. All it took was a link to access this information.

This is reported by Ars Technica, citing research by the cybersecurity company Claroc.

Grok Build works on the principle of “vibecoding”: a person describes what application they want to get, and AI writes the code for them. The result can be shared with others using a link. It is these links that have become a problem.

Claroc researchers found over 43,000 Grok Build application addresses. Of these, over 35,000 were open to outsiders.

Experts found, in particular:

  • 24 applications from SpaceXSI itself, the merged company of SpaceX and xAI. They found a confidential onboarding plan for a new employee, information about 594 customer accounts, internal addresses of operating systems, and advertising offers.

  • Internal Tesla statistics with the names of 22 service center technicians, data on their productivity and bonuses.

  • Defense contractor documents, including a list of 108 people who worked on a contract with the US Navy, and production instructions from another company.

  • Data from thousands of employees and clients of other organizations, including a list of approximately 3,500 employees of a healthcare company and information about the budgets of potential home buyers.

The problem is the way Grok Build creates links. Every published application gets an address with random words before grok.me. It was assumed that an outsider would not be able to guess such an address.

But when someone opens the Grok Build app, their web address can be entered into special databases that collect information about DNS queries. The data from such databases is publicly available, so the researchers were able to find thousands of links that should have been difficult to guess.

Claroc also tried to publish a test app created with the AI ​​assistant. It turned out that Grok Build, when publishing the app, offered access to anyone with the link. To restrict it, the user had to change the settings themselves.

The researchers also note that a similar risk exists in other AI app development tools, including Lovable, Bolt, and Replit, which also use separate web addresses for their projects.

Therefore, Grok Build users who work with internal documents or personal data are advised to check their application settings. Instead of "Anyone with the link", it is better to choose "Invited users" or "Only you".

Previously, dev.ua wrote about a study that showed that many applications created using "vibecoding" have security problems. Experts checked thousands of programs created using vibecoding services Lovable, Replit, and others. It turned out that about 40% of the studied applications disclosed confidential data, including medical and financial information or customer contacts. In some cases, it was also enough for an outsider to simply know the address of the application.

How not to find a security hole? Ukrainian developer names top 5 typical mistakes of beginners
How not to find a security hole? Ukrainian developer names top 5 typical mistakes of beginners
On the topic
How not to find a security hole? Ukrainian developer names top 5 typical mistakes of beginners
OpenAI and Anthropic API keys leaked on OpenCode. Attackers could use paid AI models for other people's money
OpenAI and Anthropic API keys leaked on OpenCode. Attackers could use paid AI models for other people's money
On the topic
OpenAI and Anthropic API keys leaked on OpenCode. Attackers could use paid AI models for other people's money
Read the country's main IT news in our Telegram
Read the country's main IT news in our Telegram
On the topic
Read the country's main IT news in our Telegram
Also Read
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
4 comments
Як ламають відео-ігри та викладають їхні піратські копії? Розповідає програміст
Як ламають відео-ігри та викладають їхні піратські копії? Розповідає програміст
Як ламають відео-ігри та викладають їхні піратські копії? Розповідає програміст
Що, юний хакер, тобі цікаво, які ігри ще досі не крякнули? Тоді мерщій читай цю статтю. Нижче ми розглянемо, які технології використовуються для захисту ігор від злому. Також не пройдемо повз рекордсменів. Дізнаємося про рекордний час, за який вдалося зламати гру. Та розглянемо справжніх «міцних горішків».
9
Шахраї грабують українців від імені «Дії» та Зеленського: перелік сайтів
Шахраї грабують українців від імені «Дії» та Зеленського: перелік сайтів
Шахраї грабують українців від імені «Дії» та Зеленського: перелік сайтів
10
Кіберполіцейські розробили онлайн-гру, що допоможе дітям виробити навички безпечної поведінки в інтернеті: як скачати
Кіберполіцейські розробили онлайн-гру, що допоможе дітям виробити навички безпечної поведінки в інтернеті: як скачати
Кіберполіцейські розробили онлайн-гру, що допоможе дітям виробити навички безпечної поведінки в інтернеті: як скачати
6

Have important news to share? Message our Telegram bot

Key events and useful links in our Telegram channel

Discussion
No comments yet.